Con inuous In eg a ion o Ne wo ks Suppo ing
Low-La ency Using Hyb id Ne wo k Emula ion
Flo ian Wiedne , Dominik K eu ze , Jonas And e, Geo g Ca le
Depa men o Compu e Enginee ing
Technical Uni e si y o Munich
Ga ching by Munich, Ge many
{ lo ian.wiedne , dominik.k eu ze , jonas.and e, ca le}@ um.de
Abs ac —Enabling con inuous in eg a ion (CI) cycles o ne -
wo k p o ocols and se ices poses a signi ican challenge due
o he necessi y o building comple e and complex ne wo ks o
es ing and e i ica ion. This p ocess demands obus simula ion,
emula ion, o a a ie y o ha dwa e esou ces.
Fo non-la ency, h oughpu -sensi i e se ices ha deal wi h
bes -e o a ic, ools like Minine o ns3 can be u ilized e ec-
i ely. Howe e , la ency-sensi i e applica ions equi e e i ica ion
in ci cums ances ha closely esemble eal-wo ld en i onmen s.
Tools such as ns3 ope a e a an abs ac ion le el ha is oo high
o accu a ely e lec eali y, while o iginal Minine , by elying on
i ual E he ne pai s, ends o lack ealis ic la ency.
To add ess his challenge, we p opose using Minine as a
s anda d and ep oducible API, enhanced wi h Single-Roo -IO-
Vi ualiza ion (SR-IOV)-based connec ions when s abili y and
lowe la encies a e pa amoun . This app oach enables us o
es and e i y wo king con igu a ions in a s aigh o wa d, non-
ha dwa e-suppo ed en i onmen , allowing he inal s ages o ou
CI p ocess o p og ess owa ds mo e s able p oduc s wi hou he
necessi y o adap sc ip s o con igu a ions due o eusing he
same API o di e en unde lying echnologies. Ou indings
demons a e ha inco po a ing SR-IOV in o ne wo k emula ion
can po en ially double he usable bandwid h and signi ican ly
enhance bo h s abili y and la ency.
Index Te ms—low-la ency, h oughpu , single- oo -io-
i ualiza ion, emula ion
2025 36 h In e na ional Tele a ic Cong ess (ITC 36)
I. INTRODUCTION
Con inuous In eg a ion (CI) o e s signi ican ad an ages
owa ds de eloping, ex ending, and imp o ing ools using es -
ing and e i ica ion o iden i y p oblems ea ly [1]. Adap ing
ne wo k p o ocols and se ices is e en mo e challenging as
ailu es can impac o he de ices and se ices in he ne wo k.
The complex na u e o ne wo ks causes challenges when
analyzing p o ocols and se ices in CIs as he in e -wo king o
componen s such as ou e s and i ewalls mus be conside ed.
This complexi y esul s in a limi ed es ing co e age pe o med
in oday’s ne wo ks. Digi al Twin Ne wo k (DTN) p o ides he
oppo uni y o u ilize a digi al eplica e o he ac ual ne wo k
o pe o m es s and e i ica ion wi hou dis u bing he o iginal
ne wo k, simila o o he esea ch a eas [2].
Se e al ools p o ide a DTN based on ei he ma hema ical
models [3]–[6], simula ions [7], [8], o emula ions [9], [10].
DTNs based on ma hema ical models p o ide, simila o
simula ions, an abs ac ion o he ne wo k add essing speci ic
aspec s o he eal ne wo k based on hei use case wi hou
he abili y o es o e i y applica ion code di ec ly. Howe e ,
ne wo k emula ions such as Minine using Linux on-boa d
ools allow o apidly build an en i e ne wo k o es and
e i y se ices and p o ocols au oma ically wi hou a physical
ne wo k [9]. Especially owa ds la ency, Wiedne e . al. [11]
compa ed in a p e ious wo k Minine o a i ual machine
(VM)-based app oach wi h links based on Single-Roo -IO-
Vi ualiza ion (SR-IOV), which ex ends he PCI speci ica ion
enabling spli ing one NIC in o mul iple ligh weigh PCI
de ices, showing ha he esul s a e nei he ealis ic no s able.
This pape analyze a hyb id ne wo k emula ion ool based
on Minine wi h SR-IOV and VMs. We aim o pe o m a
con olled compa ison using an ex e nal measu emen in as-
uc u e o p o ide insigh s owa ds analyzing la ency-s able
and high h oughpu en i onmen s. We p o ide:
1) A ex ended Minine amewo k wi h SR-IOV and VMs.
2) An ex ended pe o mance analysis.
3) Pe o mance enhancemen s o sensi i e ne wo k es s.
4) Recommenda ions on using he di e en ypes in com-
mon scena ios such as CI o DTN.
The pape is s uc u ed as ollows: Sec ion II p o ides back-
g ound in o ma ion and he cu en s a e o he a , ollowed by
challenges when using CIs o low-la ency ne wo k se ices
in Sec ion III. We ou line ou me hodology o ex ended he
Minine amewo k in Sec ion IV. Followed by he measu e-
men se up in Sec ion V and he e alua ion in Sec ion VI.
We inalize he pape wi h limi a ions, ecommenda ions,
ep oducibili y, and conclusion in Sec ions VII o X.
II. BACKGROUND AND RELATED WORK
This sec ion p esen s an analysis o ele an li e a u e in he
ields o ne wo k simula ion and emula ion, DTN, and CI.
A. Ne wo k Simula ion and Emula ion
One impo an aspec o ne wo k simula ion and emula ion
is he oppo uni y o es and e i y ne wo k beha io in
a closed, con olled en i onmen , allowing apid es ing a
di e en abs ac ion le els. I gi es he lexibili y o ebuilding
a bi a y ne wo k opologies, including complex ne wo k com-
ponen s and hei in e -wo king [12]. Simula ion uses ma he-
ma ical models, whe eas emula ion uses abs ac ion ools o
ep esen he physical ne wo k.
Se e al ools p o iding gene al-pu pose ne wo k simula ion
and emula ion acili ies a e a ailable, such as, e.g., OMNe ++,
NS3, o Minine [13]–[15] p o iding a ich ange o ea u es.
Wiedne e al. [11] show ha Minine is no designed o
ealis ic ail-la ency and h oughpu . They p opose using ca-
bled SR-IOV in e aces o eplace i ual links wi h highly
op imized VMs as ne wo k nodes and show ha hey can
achie e ealis ic low and s able la encies bu did no in eg a e
i in an au oma able amewo k needed in CI. Fu he mo e,
simula ion ools such as OMNe ++ [13] and NS3 [14] a e
based on abs ac ma hema ical models o he eal wo ld o
desc ibe he ne wo k in disc e e e en s.
Minine [15] uses Linux on-boa d ea u es and a Py hon
API o buil a ne wo k by isola ing di e en nodes using Linux
namespaces, and using i ual E he ne pai s ( e h) as links.
This emula ion a chi ec u e allows using Linux s anda d ne -
wo k applica ions in eal- ime. Fu he mo e, Yan and Jin [16]
a e o e coming he challenge o educed bandwid h capa-
bili ies in Minine when he expe imen load is highe han
he hos ’s physical esou ces by adding a i ual ime o he
nodes imp o ing he esul s’ ealism. Howe e , i imp o es
ail-la ency a o e eali y as i does no conside he ime-
wise co ela ion o e en s on di e en nodes. Con aine ne , a
Minine o k, p o ides Docke con aine s as ne wo k nodes
and ea u es such as CPU o memo y isola ion using he
Minine API [17]. Wi h Con aine ne 2.0 Peus e e . al. [18]
added suppo o in eg a ing VMs in o Minine , using e hs
as connec ions based on Con aine ne s API. As VMs allow o
use applica ions equi ing ke nel access, ou ocus in his pape
lies on ex ending he app oach o Con aine ne 2.0 by adding
VMs and ha dwa e-based connec ions o euse he exis ing
API o ail-la ency analysis.
B. Single-Roo Inpu /Ou pu Vi ualiza ion
One ex ension o u ilize he lexibili y o p o iding a bi a y
opologies wi hou he equi emen o cabling and suppo ing
p ocessing packe s on eal-ha dwa e o imp o e pe o mance
is SR-IOV. SR-IOV ex ends he PCI speci ica ion, enabling
he spli ing o one physical unc ion (PF) in o mul iple
i ual unc ions (VF)—ligh weigh PCI de ices. Each VF
has a sepa a e send and ecei e queue dis inc om he PF
queues o educe o e head and in ol emen o he OS in
packe p ocessing [19]. Liu [20] and Dong e al. [19] ha e
bo h demons a ed ha SR-IOV is imp o ing bo h bandwid h
u iliza ion and la ency in compa ison o d i e s such as i io
o dis ibu e he packe s om one physical link in o mul iple
i ual de ices. Using SR-IOV, i is possible o achie e line
a e in each node wi hou he need o pass h ough one physical
ne wo k in e ace ca d (NIC) pe node. Simila o [11] a e we
u ilizing SR-IOV in combina ion wi h VLAN-IDs o p o ide
a ic o mul iple link connec ions o e one cable.
C. Digi al Twin Ne wo k
Ne wo k emula ion can be he basis o DTNs p o iding
a i ual win o he physical in as uc u e o analyze and
op imize ope a ions and beha io s [2]. DTNs can be u he
buil using g aph-neu al ne wo ks [5], simula ion, o AI [4]
Tang e . al. [21] ou line in hei esea ch he need o delay-
sensi i e a ic wi hin DTNs as se e al po en ially analyzed
se ice-le el ag eemen s equi e his. Ou app oach allows o
use Minine as la ency-sensi i e DTNs.
D. Con inuous In eg a ion
CI is he cen al concep o allow au oma ic es ing and
e i ica ion o sys ems be o e deploymen coming om he
Ex eme P og amming de elopmen p ocess [1]. CI is an
au oma ed p ocess execu ed on each commi in a so wa e
eposi o y, p o iding a di ec eedback loop o he de elope s.
I imp o es code quali y h ough es ing and en o ces code
s yling [22]. CI includes au oma ic build and es so wa e
be o e new ea u es o bug ixes a e deployed in o p oduc ion,
educing he human e o in es ing he so wa e and educe
he e o a e o deployed so wa e.
Vucnik e al. [23] desc ibe he use o CI o wi eless
communica ion as challenging due o hei complex na u e. A
unique sys em is needed in eg a ing web echnologies and OS
i ualiza ion in exis ing 5G es beds o enable CI o wi eless
communica ion. DigSiNe [24] is an example o a con aine -
based DTN used o CI in a es bed in as uc u e o p edic
ailu es in ne wo k con igu a ions. Ou pape suppo s he idea
o ne wo k CI wi h a amewo k o au oma ed es ing o low-
la ency and high h oughpu se ices.
E. Low-la ency in Vi ualized Sys ems
Vi ualiza ion enables esou ce sha ing and lexible, on-
demand p o isioning o esou ces, especially in he con ex o
DTN and CI his is a equi emen due o hei as -changing
and adap ing na u e. Two commonly used a chi ec u es o
i ualiza ion a e hype iso - and con aine -based. VMs a e
Hype iso -based i ualiza ions isola ing he comple e OS,
including he ke nel. Whe eas con aine s a e called ligh weigh
i ualiza ions using a sha ed ke nel [25]. Se e al s udies
such as [20], [26]–[29] analyzed la ency in bo h con aine s
and VMs p o iding aluable insigh s in o he di e ences
be ween bo h echniques. Daichend e . al. [30] analyze he
in luence o mul iple con aine s on each o he , showing ha
he in luence o he nodes, especially owa ds la ency, is
signi ican . Wiedne e al. [11] showed ha using se e al OS-
le el op imized VMs and SR-IOV imp o es he end- o-end
la ency and s abili y signi ican ly [11]. We ex end Minine wi h
echnologies p oposed by [11] o imp o e la ency s abili y.
III. CONTINUOUS INTEGRATION AND LOW-LATENCY OF
NETWORK SERVICES
CI p o ides a long lis o ad an ages owa ds con inuous
ea u e de elopmen in so wa e sys ems [1]. Howe e , when
analyzing ne wo ks and hei se ices, and he communica ion
s uc u e, i is a challenge o in eg a e a CI in a nea - o- eali y
en i onmen [17]. Emula ions allowing apid p o o yping in
en i e ne wo ks a e needed o analyze ne wo k se ices and
p o ocols in a DTN enabling he use o Linux ne wo king
so wa e di ec ly. This design imp o es he possibili y o
0 50 90 99 99.9 99.99 99.999 99.9999
100
101
102
103
104
105
106
107
Pe cen iles [%]
La ency [µs]
50 300 500
700 900 1000
Figu e 1: La ency HDR diag am ac oss a ious a es
(kpacke s/s) send h ough o iginal Minine .
es ing se ices ha equi e di e en pa s dis ibu ed wi hin
he ne wo k. To allow execu ing so wa e equi ing ke nel
access, Peus e e al. [18] ex end Minine wi h he oppo uni y
o use VMs as Minine nodes. Minine ’s lexibili y allows o
he apid building o gene al-pu pose ne wo ks au oma ically.
To illus a e cu en la ency and h oughpu issues, we pe -
o m measu emen s wi h cons an -bi a e (CBR) a ic om
an ex e nal load gene a o wi h a packe size o 84 B and
a se en nodes line- opology wi h nodes di ec ly connec ed
o each o he wi hou swi ches in Minine using a anilla
Debian Bookwo m OS. The maximum s able packe a e in
ou se up on Minine is 270 kpacke s/s. Wi h a es below
270 kpacke s/s, we achie e a s able packe a e p ocessed
h ough Minine . Figu e 1 shows he la ency esul s a di e en
a es om 50 kpacke s/sun il 1000 kpacke s/sin a high
dynamic ange (HDR) diag am wi h bo h axes loga i hmic.
The x-axis p esen s he pe cen iles o he la ency shown on
he y-axis. Simila o p e ious esul s such as [11] is he
a e age and ail-la ency eaching o e 80 ms in compa ison
o expe imen s using eal cabled connec ions. Addi ionally we
obse e 20 % packe loss wi h 500 kpacke s/sin his scena io
and as he packe loss inc eases, he wo s -case la ency s ays
mos ly s able due o d opped packe s educing he load on
he memo y-based links. This al eady shows he impo ance
o imp o ing he po en ials o Minine .
This is especially p oblema ic when analyzing la ency-
sensi i e ne wo king se ices and applica ions as he la ency
alues a e mainly caused by high load, no he analyzed appli-
ca ion. When ne wo k se ices ha e s ic equi emen s on he
a ic such as low- ail-la encies o de e minis ic bandwid h,
hen i ualiza ion is a challenge as Gallenm¨
ulle e al. [27]
ou line. They show ha mos la ency ou lie s in i ualized
sys ems a e caused by in e up s due o he highe cos o
in e up ing i ualiza ion. Typically, such sys ems equi e a
s ic la ency un il he 99.9999 h pe cen ile; a e ou lie s in
highe pe cen iles a e no conside ed [27]. Especially when
using namespaces o isola ion he in luence o he co e on
which he p ocess is unning and he in e up a ini y is
signi ican compa ed o ba e-me al sys ems [30].
S able la encies a e essen ial o model he di e ences be-
ween emula ed and eal ne wo ks, analyze he la encies, and
con i m ha he p ocessing delay in combina ion wi h he
ne wo k a ic, o example, holds he equi emen s de ined
o he p o ocol o applica ion. Based on he esul s shown
in Figu e 1, he pape aims o analyze he po en ials o
using a de ined API such as Minine and he po en ial o
apidly build ne wo ks and allow apid p o o yping and es ing
o ne wo k applica ions in combina ion wi h s a e-o - he-a
op imiza ions o allow analyzing low-la ency, de e minis ic
ne wo k applica ions, se ices, and p o ocols. Ou mo i a-
ion example in Figu e 1 shows ha cu en ly, he la ency
in Minine is eaching much ea lie han ealis ic ne wo k
appliances an o e load and he e o e ha ing highe packe loss
and a high gap be ween lowe and highe pe cen ile la encies
making i un ealis ic o be used in CI au oma ion o low-
la ency ne wo k se ices such as Time Sensi i e Ne wo king
algo i hms o be es ed. Iden i ying i i is possible o achie e
bo h lexibili y and low la encies is he main a ge o his
pape , concluding ha inding a way o u ilize a s anda d API
o suppo easy in eg a ion o CI au oma ion and op imizing
owa ds de e minis ic and eliable la ency is essen ial.
IV. METHODOLOGY
We p esen ou me hodology o ex end Minine owa ds
mo e eliable bu s ill apid p o o yping when low-la ency o
de e minis ic bandwid h a e equi ed. We ex end Minine wi h
VMs simila as done p e iously in Con aine ne 2.0 [18].
Addi ionally, we u ilize SR-IOV as ha dwa e connec ions
based on Wiedne e al. [11] o p o ide hyb id emula ion o
opologies. Fu he mo e, we use a small se o op imiza ions
on he hype iso as hei usage depend on he con ol o e
he hype iso and co esponding boo pa ame e a e hose no
in eg a ed in o he ex ended Minine e sion.
Wi h ex ending Minine , we aim o p o ide a ealis ic
CI en i onmen o low-la ency o de e minis ic bandwid h
ne wo k se ices and p o ocols. Addi ionally, we show ha
using ou app oach, we achie e a lexible and scalable solu-
ion o emula e complex opologies and p o ide pe o mance
benchma ks o ne wo k se ices as pa o DTNs o CIs
ha can be ex ended owa ds a de elopmen o measu emen
pla o m.
Using VMs, he ange o a ailable ne wo k se ices and
p o ocols is ex ended as ke nel access is a ailable as equi ed
in some applica ions. Ca e ully op imized VMs show simila
pe o mance owa ds bandwid h and ail-la ency as Linux
con aine s bu equi e mo e o e head in esou ce usage as
examined by Wiedne e al [26]. Figu e 2 shows all ou a i-
an s analyzed in his pape . The abs ac a chi ec u e neglec s
he Minine py hon API and he hype iso . In Figu e 2a, he
adi ional Minine app oach wi h namespaces and e h links
is p esen ed as baseline o ou ex ended Minine amewo k.
Ou me hodology includes he ollowing a ia ions: combining
VMs wi h e h based links o eplacing he links wi h SR-IOV-
based VFs using Vi ual LAN (VLAN) IDs in bo h a ian s
as shown in Figu es 2b and 2d o dis inguish he links on he
Node 2Node 1 Node 3
NIC A NIC B
(a) Namespace wi h e h
Node 2Node 1 Node 3
NIC A NIC B
402 403 402 403
(b) Namespace wi h SR-IOV VF
VM 2VM 1 VM 3
NIC A NIC B
(c) VM wi h e h
VM 2VM 1 VM 3
NIC A NIC B
402 403 402 403
(d) VM wi h SR-IOV VF
Figu e 2: Node-link ype combina ions analyzed—link label
ep esen ing VLAN IDs.
sha ed cable. Using his app oach, we can use swi ches wi hin
Minine , and he connec ions a e wo king o bo h mul icas
and unicas connec ions as explained by Wiedne e al. [11].
To abs ac he node ype om he use in ou app oach, we
added a new node ype VMNodeMixin as basis ex ended o-
wa ds swi ches o hos s. This class ansla es be ween QEMU,
a gene ic engine i ualizing machines, and Minine API.
In ou me hodology, QEMU, combined wi h ke nel i ual
machines (KVM), u ilized as hype iso o VMs. This allow
us o use he same Minine sc ip s o VMs and namespace
nodes as well as he lexible change be ween node ypes. To
enable adding in e aces in Minine , hey a e added a e he
node is boo ed, using a ho -pluggable PCI Exp ess o PCI
b idge. In e aces a e ei he added di ec ly in pass h ough
mode o using MACVTAP de ices as in e media e i di ec
pass h oughs a e no suppo ed. VM-based nodes expose mul-
iple con igu a ion op ions o le he use con igu e he numbe
o CPUs, memo y, he pa h o he ke nel, ini d and ile-sys em
images, and CPU pinning o allow op imiza ions [27]. The
communica ion be ween Minine and he VM node is sol ed
using a Unix socke passed o he gues . A Minine shell is
spawned inside he VM connec ed o e he i ual console o
he Minine p ocess on he hype iso . Adding a new node
ype makes he usage simple and ex ensible.
To imp o e he pe o mance, we wan o suppo ha dwa e
in e aces, in ou case, SR-IOV VFs, by de ining he in e ace
on which VFs a e a ailable. The co esponding numbe o
VFs pe in e ace and pai -wise VLAN IDs mus be con ig-
u ed be o e. This app oach educes he needed adop ions in
Minine when he ha dwa e changes, as he con igu a ion mask
including spawning o VFs is endo -speci ic. Ha dwa e links
a e added as a new link ype HwPai expec ing a pai o
al eady exis ing in e aces connec ed di ec ly o , o example,
using SR-IOV and VLAN-IDs o sepa a ion. To simpli y his
p ocess, a global lis can be p o ided o Minine wi h pai s
o in e aces each ep esen ing a possible di ec connec ion.
Access o he code is desc ibed in Sec ion IX.
Ou me hodology can be used wi h mos Minine o ks.
We implemen ed ou p o o ype using he o iginal Minine
as he ea u e se is mo e condensed han o ks such as
Table I: La ency op imized boo pa ame e s o Hos OS.
Pa ame e Value Desc ip ion
i qa ini y 0 In e up s on speci ic co e
idle poll Poll mode when co e idle
sc eliable Rely on TSC wi hou check
mce igno e ce Igno e co ec ed e o s
audi 0 Disable audi messages
nmi wa chdog 0 Disable NMI wa chdog
skew ick 1 No simul aneous icks o locks
noso lookup Disables logging o back aces
nosm Disables hype h eading
LoadGen DuT
Minine
Times ampe
▶
◀
▶
◀
▶
◀
▲ ▲
Figu e 3: Measu emen -se up de i ed om Wiedne e al. [26].
Con aine ne . The e o e, adding ex ensions such as SR-IOV
and VMs is mo e s aigh o wa d. Mo eo e , mixing di e en
ypes o nodes and links in one expe imen is possible by
using di e en node and link classes o he de ini ion o he
opology. The ollowing sec ion ou lines how ou measu emen
se up enables o p ecisely e alua e he pe o mance o Minine
and ou ex ensions and educes he in luence o he load
gene a ion and measu emen on o he esul s.
V. MEASUREMENT SETUP
As ou lined in Sec ion II, in e up s, among o he ea u es
like he sleep s a e o he CPUs and memo y copies, a e
he highes in luence ac o on o wa ding la ency wi hin
one machine. Wi h building an en i e ne wo k opology on
one machine, he in luence is e en highe . In se e al s udies
such as [26], [27], [31] ne wo k la ency op imiza ions a e
explained, and he lowes la encies a e achie ed when using a
combina ion o OS-le el op imiza ions, use -space ne wo king,
and di ec access o he NIC. In ou se up, we use a base-
line o op imiza ions such as isola ion echniques, disabling
wa chdogs, and a lis o addi ional op imiza ions o educe
he impac o audi ion and sleep s a es o he CPU on ou
measu emen s [26]. The selec ed pa ame e s a e shown in
Table I. These op imiza ions include disabling ene gy-sa ing
mechanisms, mo ing he in e up a ini y o a speci ic co e,
and u ning o audi messages. In e up s a ini y can be se
o a speci ic co e, and logging o back aces can be educed o
imp o e la ency. Fo he VM, we analyze wo di e en modes,
a non-op imized wi h only he baseline op imiza ions and an
op imized e sion wi h pinned and isola ed co es o educe he
in luence o in e up s on he la ency o educing he o e head
caused by using ull i ualiza ion.
To allow o p ecise measu emen s wi hou in luencing he
De ice-unde -Tes (DuT) unning Minine wi h he measu e-
men p ocess i sel , we use a h ee ha dwa e-machine-based
se up as p oposed by Wiedne e al. [26]. Figu e 3 depic s he
s uc u e based on a load-gene a o (LoadGen), a imes amping
machine ( imes ampe ), and he DuT unning he so wa e o
analyze. The LoadGen uses an In el Xeon Sil e 4116 CPU,
192 GB RAM, and a dual-po In el 82599ES 10-Gigabi SFP+
NIC connec ed o he DuT using op ical ibe s. We use he sec-
ond machine as imes ampe connec ed o he ibe s be ween
DuT and LoadGen wi h passi e op ical e minal access poin s
(TAPs) o ensu e high p ecision measu emen s pe packe a
line a e. These TAPs add he same cons an delay on bo h
sides, which can be neglec ed as i is au oma ically emo ed
in he la ency calcula ion consis ing o he ime be ween he
wo measu ed imes amps. The imes ampe is equipped wi h
an AMD EPYC 7542 32-Co e P ocesso , 128 Gbi o RAM
and a dual-po In el E810-XXV 25 Gbi /sNIC lashed o
10 Gbi /s. The ha dwa e imes amping o he In el E810 NIC
o e s a p ecision o 1.25 ns [32].
The DuT is equipped wi h an AMD EPYC 7551P 32-Co e
P ocesso , 128 GB RAM, and 2 ×In el X710 10GbE SFP+
NICs wi h one po o each NIC is linked o he LoadGen.
The o he wo po s a e linked oge he and used as a loop
cable o he wo se ups wi h SR-IOV-based links as shown
in Figu es 2b and 2d.
The se up shown in Figu e 3 allows a p ecise analysis
o packe p ocessing ail-la ency o ocus on compa ing he
di e en con igu a ions wi h each o he ins ead o he di e en
measu emen execu ion in luences. We use MoonSni [33] o
MoonGen on he LoadGen and imes ampe o ansmi and
eco d minimally sized packe s wi h 84 B. The size o 84 B
is he minimum packe size, allowing a la ge enough ID o
ansmi each packe as pa o he payload. I no o he wise
speci ied, we a e using CBR a ic in all expe imen s.
We u he use Debian Bookwo m 11 (ke nel 6.1) as bo h
hype iso and VM OS o u ilize s able so wa e lib a y e -
sions on a cu en ly a ailable OS e sion. Ou Minine e sion
and QEMU a e unning inside he DuT, p ocessing packe s
a i ing om he LoadGen and sending hem back o he
LoadGen a e a e sing h ough he opology deployed wi h
Minine . The QEMU API and emula o is used o execu ing
KVM VMs. Fo a deepe unde s anding o he sc ip s and code
execu ed, we edi ec eade s owa ds Sec ion IX.
VI. EVALUATION
To assess he p oposed ex ended Minine me hodology, we
i s p esen he scena io used o ou e alua ion, conduc
baseline measu emen s, and analyze his scena io’s po en ial
usable bandwid h and la ency. Following his, we compa e
an op imized e sion wi h an unop imized e sion based on
VM measu emen s. We conclude ou e alua ion by analyzing
a scena io wi h bu s iness in a ic and i s in luence on he
a ian s. The esul s ob ained a e used o o e he eade
ailo ed ecommenda ions o a ious usage scena ios.
A. Scena io
To e alua e he Minine opology, we selec a con igu a ion
ha is bo h s aigh o wa d and capable o s essing he sys em
using CBR a ic. Addi ionally, he scena io is designed o
0 20 40 60 80 100
0
200
400
600
800
Measu emen ime [s]
Packe a e [kpk /s]
e h/namespace
SR-IOV/namespace
e h/VM
SR-IOV/VM
Figu e 4: Packe a e o e ime a e passing h ough
Minine —LoadGen ansmi CBR a ic wi h 700 kpacke s/s.
illus a e he in e ac ions among mul iple nodes wi hin he
same sys em.
We u ilize a line opology comp ising se en nodes, labeled
h1 h ough h7, whe e h1 is esponsible o ecei ing a ic
om he LoadGen, and h7 ansmi s he a ic back o he
LoadGen. No u he nodes a e in ol ed, he nodes a e di ec ly
connec ed o each o he educing he complexi y o simple
unde s andabili y o he esul s.
To acili a e a ic o wa ding among he di e en nodes,
all hos s employ Linux ou ing and o wa ding capabili ies,
ou ing a ic based on hei espec i e ou ing ables o
he nex hop in line. This scena io e ec i ely s esses he
applica ion wi h high packe a es and yields aluable insigh s.
We in es iga e a selec ed ange o packe a es—50, 300,
500, 700, 900 and 1000 kpacke s/s—ac oss all ou a ian s
depic ed in Figu e 2. Ou aim is o analyze how he di e en
a ia ions espond o a ying packe a es in ela ion o la ency,
as well as o iden i y he maximum packe a e ha each
a ia ion in he same scena io can e ec i ely p ocess.
B. Baseline measu emen
To unde s and he changes iden i ied in he e alua ion o he
newly added a ian s using SR-IOV and/o VMs in Minine ,
we i s e alua e a baseline wi h he o iginal, no adap ed
Minine . Fo iden i ying he bandwid h bo leneck, he packe
a e is displayed in Figu e 4 o e he ime ha was ecei ed
a e ansmission h ough Minine wi h CBR a ic. We
selec ed 700 kpacke s/sas a e o p esen in he diag am
o e ime a e analyzing he da a in Figu e 4 as blue line.
700 kpacke s/sis highe han Minine can handle he a ic
in eal- ime showing clea ly he achie able packe a e. The
maximum packe a e ha can be handled wi hou packe loss
is 270 kpacke s/s. We aim o o e come his limi a ion by
adding SR-IOV and VMs o enable highe packe a es wi hou
packe loss.
Ano he impo an me ic, as shown in ou mo i a ing
example in Figu e 1, is la ency, especially ail-la ency. We
p esen an HDR diag am wi h he la encies o all analyzed
a ic a es in Figu e 1 showing ha he di e ences be ween
packe a es a e signi ican , anging in wo s case om 688 µs
un il 4 s. The la encies o 50 kpacke s/sa e showing a mean
o 68 µs ising owa ds mo e han 303 µsin wo s -case wi h
a s anda d de ia ion (STD) o 18.7µsand a wo s -case ji e
o 44 µsshowing low-la encies be o e packe loss occu s wi h
a high ji e de i a ion. Las ly, he packe a es highe han
270 kpacke s/swi h signi ican packe loss ha e a ail la ency
abo e 70 ms due o memo y o e head. Resul ing in gene al
la ency luc ua ions be o e packe loss occu s, la encies wi h
a es causing packe loss, and, he e o e, changed ne wo k
beha io a e beha ing in gene al as expec ed. In summa y,
when de e minis ic la ency and high h oughpu a e equi ed,
we showed ha ano he apid p o o yping and CI ool is
needed compa ed o adi ional Minine .
C. Th oughpu
One o he a ge s o his pape is o highligh he di e ences
owa ds de e minis ic and high bandwid h u iliza ion be ween
he o iginal and ex ended Minine e sion. The p e ious
subsec ion shows ha he o iginal Minine e sion has a
bandwid h bo leneck a 270 kpacke s/s. We pe o m he same
compa isons o all o he a ian s o iden i y hei bo leneck
bandwid h and h oughpu possibili ies. The ecei ed a e a e
each a ian when sending 700 kpacke s/s om he LoadGen
is shown in Figu e 4 compa ed o he o iginal Minine a ian
in blue.
When changing he node ype om namespaces owa ds
VMs while s ill u ilizing e hs as links, he maximum packe
a e achie able dec eases d ama ically o a maximum o
145 kpacke s/sins ead o 270 kpacke s/s. This dec easing
maximum loss-less a e is esul ing om he addi ional added
memo y copies and o e head due o adding a i ual in e ace
o a VM ins ead o u ilizing di ec memo y access (DMA)
and IOMMU ansla ions as possible when using a PCI de ice
ins ead. Based on hese esul s only changing he node ype
as done in Con aine ne 2.0 [18] is no achie ing ou a ge o
de e minis ic and high h oughpu .
As we added in ou app oach one addi ional change op-
po uni y, we analyze he impac o adap ing he link ype
om e h o SR-IOV while using namespaces as node ypes
o make i compa able o he o iginal Minine . Changing he
link ype om a memo y copy-based i ual in e ace o a
ha dwa e-based in e ace achie es a high inc ease o s able
h oughpu . Wi h his combina ion we can achie e a s able a e
o 680 kpacke s/sin ou measu emen s, which is an inc ease
o 410 kpacke s/s. This inc eased capabili ies clea ly show
ha e hs as Linux implemen a ion o i ual in e aces used
in Minine a e p o iding he highes bo leneck due o bo h,
he nodes and he a ic on he link, equi ing capaci y om
CPU, memo y, and sys em buses.
When u ilizing bo h VMs as nodes and SR-IOV VFs as link
endpoin s wi h DMA and di ec pass h ough, we can achie e a
e en highe h oughpu wi hou packe loss a 720 kpacke s/s
esul ing due o he highe isola ion o VMs compa ed o
namespaces. This clea ly shows ha o expe imen s equi ing
0 50 90 99 99.9 99.99 99.999 99.9999
100
101
102
103
104
105
106
107
Pe cen iles [%]
La ency [µs]
Ra e link/node
050 500 e h/namespace SR-IOV/namespace
300 700 e h/VM SR-IOV/VM
Figu e 5: La ency HDR diag am o a es wi hou packe loss
o all node-link a ian s.
high and de e minis ic as well as s able h oughpu capabili ies,
changing he link ype o a ha dwa e suppo ed, less CPU and
memo y in ense in e ace is necessa y. Adding he addi ional
o e head and highe esou ce demand o VMs in compa ison
is imp o ing he esul s sligh ly, bu he esou ce consump ion
is inc easing d ama ically in compa ison.
D. La ency
A e analyzing he maximum possible h oughpu wi hou
packe loss on each a ian , we wan o see i we can achie e a
s able and low la ency on each a ian analyzed in his pape .
Figu e 1 is showing he baseline o hese measu emen s.
We conside o ou u he analysis only hose a es wi hou
packe loss o each a ian as la encies in o e load scena ios
a e no de e minis ic as shown by Gallenm¨
ulle e al. [34].
Figu e 5 is p esen ing one HDR diag am con aining he la en-
cies o all ou a ian s o a es wi hou packe loss. When
analyzing all a ian s a 50 kpacke s/s, we see a ail-la ency
o he o iginal Minine a 303 µs(mean 68 µs, STD 18.7µs,
wo s -case ji e 44 µs), VM wi h namespaces a 9211 µs(mean
420 µs, STD 173 µs, wo s -case ji e 60 µs), VM wi h SR-IOV
1596 µs(mean 196 µs, STD 371 µs, wo s -case ji e 37 µs),
and namespaces wi h SR-IOV 455 µs(mean 148 µs, STD
93 µs, wo s -case ji e 37 µs) showing ha in he case o no
packe loss o memo y o e load, he o iginal Minine shows
a nea ly s able beha io wi h mino adap ions based on he
lowes o e head due o di ec memo y copies and ligh -weigh
isola ion, whe eas bo h a ian s wi h SR-IOV a e showing an
imp o ed ji e beha io o e se en nodes. Fo highe packe
a es, we can only analyses he wo a ian s wi h SR-IOV
based links as he a ian s wi h e hs a e causing packe loss
wi h a es highe han 270 kpacke s/s. The aise in ail-la ency
be ween he 99 h and he 99.9 h pe cen ile o VMs is simila
o esul s om p e ious wo ks such as Wiedne e al. [26]
when using pa ly op imized VMs. In his case, he VMs a e
no pinned o speci ic co es, and he co esponding co es a e
no isola ed.
Fo he highes packe a e suppo ing bo h namespaces
and VMs in combina ion wi h SR-IOV (500 kpacke s/s), he
0 50 90 99 99.9 99.99 99.999 99.9999
100
101
102
103
104
105
106
107
Pe cen iles [%]
La ency [µs]
300 500 700
Figu e 6: La ency HDR diag am on selec ed a es
(kpacke s/s) wi h SR-IOV VFs and op imized VMs
la ency o he namespace is ising owa ds 1399 µs(mean
271 µs, STD 108 µs, wo s -case ji e 19 µs) and o VMs
owa ds 23 399 µs(mean 228 µs, STD 341 µs, wo s -case ji e
741 µs). The educed ji e and simila s abili y in la ency a e
caused by he change om in e up o polling-based mode
in he Linux NAPI d i e . As expec ed, he la ency is ising
o bo h a ian s bu showing ha we can achie e a la ency
dis ibu ion simila o p e ious wo ks such as [11]. This shows
ha when highe packe a es and s able la ency a e equi ed,
namespaces wi h SR-IOV a e he bes choice, whe eas VMs
equi e u he op imiza ions o e ie e simila la ency anges.
E. Op imiza ions
Op imizing he VMs addi ional wi h pinning he i ual
co es o he VM o physical co es o he hype iso on he
NUMA node he in e ace p o iding he VFs o his node
is a ached o. Mo eo e , we isola e hose co es om he
hos ke nel o educe he impac o in e up s on o o wa ding
la ency [27]. The downside o pinning VMs o hos co es is
ha no load-balancing o o e -p o isioning o he speci ied
co es is possible esul ing in inc eased esou ce consump ion
and dec eased lexibili y. When analyzing he op imized VM
in compa ison o he non-op imized VM wi h using e h
in e aces, he op imiza ion has no in luence on he maximum
achie able packe a e due o he needed ke nel ope a ions
which a e mo ed o o he co es. Fu he mo e, he maxi-
mum achie able h oughpu using SR-IOV links on op imized
VMs in ou se en-node scena io is 850 kpacke s/scompa ed
o 720 kpacke s/swi hou op imiza ions. This imp o emen
means op imizing he VMs and isola ing he co es is impo an
o achie e high h oughpu wi h Minine , VM nodes, and SR-
IOV-based links i needed.
The imp o emen in ail-la ency shown in Figu e 6 o he
h ee highes measu ed and achie able a es is simila o he
inc ease o he maximum achie able loss-less packe a e.
Fo 500 kpacke s/sa e we achie ing a wo s -case la ency
o 4667 µs(mean 262 µs, STD 71 µs, wo s -case ji e 14 µs)
which a e highe la encies han he a ian wi h namespaces
due o addi ional o e head bu he lowes s a is ical de i a ions
0 50 90 99 99.9 99.99 99.999 99.9999
100
101
102
103
104
105
106
107
Pe cen iles [%]
La ency [µs]
SR-IOV/CBR
SR-IOV/bu s y
e h/bu s y
Figu e 7: La ency HDR diag am compa ing CBR a ic s.
bu s s o 100 packe s wi h namespaces—Ra e 300 kpacke s/s
o CBR and a ic wi h bu s s
showing s able beha io o he la ency and simila beha io
o all a es. In ou discussion on VM op imiza ions, we
demons a ed h ough ou e alua ions ha u ilizing SR-IOV o
ha dwa e in e aces ins ead o e h in Minine enhances bo h
h oughpu and la ency pe o mance. While VMs gene ally do
no in oduce signi ican o e head, hey migh no yield he
highes e enue. The e o e, i is ad isable o use his esou ce
o e head when an en i e ope a ing sys em is necessa y.
F. Bu s y T a ic
Mo eo e , in addi ion o ou analysis o CBR a ic, we
analyzed he beha io o he di e en a ian s wi h bu s y
a ic anging be ween 50 o 200 packe s pe bu s . We
wan o analyze whe he he p e ious esul s hold wi h bu s y
a ic. We ha e analyzed o namespaces he la ency o
300 kpacke s/sin Figu e 7 in compa ison o SR-IOV wi h
CBR a ic. Due o he p ocessing o packe s in a bu s ,
his a e does no cause packe loss o he o iginal Minine
compa ed o CBR a ic be o e. Fo SR-IOV wi h namespaces
is he ail-la ency a 1040 µswi h bu s y a ic which is sligh ly
lowe han o CBR a ic, whe eas as expec ed he mean ises
signi ican ly o 223 µs(STD 36 µs, wo s -case ji e 59 µs) as
shown in Figu e 7.
In Figu e 8 is he same showing o VMs wi h a a e o
700 kpacke s/scompa ing be ween non-op imized and op i-
mized VM wi h bu s y and CBR a ic. Fo he non-op imized
VM, sligh a ia ions a e isible in a small ange. In con as ,
he op imized VM di e s only in ha he jump o highe
la encies a e he 99.9 h pe cen ile is sligh ly highe o bu s y
a ic. This analysis shows ha wi h ou ex ended Minine ,
we a e able o handle bu s y a ic a a high h oughpu a e,
especially when using SR-IOV in combina ion wi h op imized
VMs o namespaces.
VII. LIMITATIONS
In addi ion o he bene i s we show in ou measu emen s,
he esul s a e limi ed o simple packe p ocessing and a sim-
ple opology. Adding compu a ions o he packe p ocessing
0 50 90 99 99.9 99.99 99.999 99.9999
100
101
102
103
104
105
106
107
Pe cen iles [%]
La ency [µs]
VM/CBR VM/bu s y
VM op imized/CBR VM op imized/bu s y
Figu e 8: La ency HDR diag am compa ing CBR a ic s.
bu s s o 100 packe s wi h VMs—Ra e 700 kpacke s/s o
CBR and a ic wi h bu s s
educes he a ailable CPU ime and inc eases he la ency,
whe eas he h oughpu migh no e en be a ec ed. We can
in e p e he analyzed highes achie able packe a e as uppe
bound o po en ial achie able packe a es. Simila , ou ana-
lyzed simple line opology shows he pe o mance di e ences,
which can widely di e in mo e complex scena ios, bu he
simple opology al eady shows he po en ial bene i s o mo e
complex opologies due o signi ican ly lowe la encies.
Mo eo e , ou measu emen se up is used o educe he
measu emen impac o an ex e nal load gene a o o e ie e
high and cons an bi a es. Due o limi ed ime and space, we
lea e an analysis o he impac o packe gene a ion wi hin
Minine on ou esul s o u u e wo k.
Fu he mo e, using SR-IOV o p o ide links in Minine
limi s he numbe o a ailable links o he numbe o VFs he
used NIC allows. The ha dwa e en o ces his limi a ion due
o he limi ed numbe o egis e s o he di e en sepa a ed
ea u es o VFs wi h he same PF. This numbe can be
inc eased by using mul iple pai s o NICs.
VIII. RECOMMENDATIONS
Depending on he speci ic equi emen s o he expe imen ,
CI, o DTN conduc ed wi h ou ex ended e sion o Minine ,
one o he ou a ian s illus a ed in Figu e 2 can be chosen.
I esou ce cons ain s a e a c i ical conce n and nei he
h oughpu no la ency needs o be p io i ized in he mea-
su emen s, we ecommend using he o iginal Minine o ou
ex ended a ian ha employs namespaces and e h-based
links. These op ions ha e he lowes impac on esou ces,
hough hey come wi h a sligh ade-o in pe o mance
ega ding packe loss.
In scena ios whe e an en i e ope a ing sys em is necessa y
on each node, and h oughpu o la ency conside a ions a e
no applicable, he bes app oach is o u ilize VMs as nodes
along wi h e h links. While his op ion may esul in he
wo s o e all pe o mance, i does no equi e addi ional NIC
and mainboa d suppo .
Table II: Summa ized Recommenda ions
Technology La ency Th oughpu Resou ces
Namespace
e h × ×× ✓✓
SR-IOV ✓✓ ✓ ✓
VM
e h ×× ×× ×
SR-IOV ✓ ✓✓ ××
When an en i e ope a ing sys em is needed on each node
along wi h high h oughpu o low la ency, using VMs in
combina ion wi h SR-IOV on Minine can o e signi ican
bene i s. I esou ces allow, we ecommend dedica ing ime o
op imize bo h he VMs and he hype iso , as his can lead o
signi ican ly imp o ed pe o mance.
Fo all o he si ua ions, we sugges using namespaces com-
bined wi h SR-IOV-based links. This con igu a ion p o ides
low and s able la ency, as well as a signi ican enhancemen
in he maximum packe a e achie able wi h ze o packe loss.
We summa ize ou ecommenda ions in Table II.
IX. REPRODUCIBILITY
We published he ex ended Minine e sion as CIMinine 1.
Fu he mo e all ins uc ions and sc ip s o ep oduce ou
esul s, addi ional measu emen s and aw da a a e a ailable
on he accompanying websi e2. The aw da a a e a ailable on
mediaTUM3.
X. CONCLUSION AND FUTURE WORK
Con inuous In eg a ion and Digi al Twins a e cu en and
u u e echnologies enabling he au oma ion o deploymen
and con igu a ion p ocesses. As ne wo ks a e complex and
mul iple de ices a e in ol ed, ealis ically, emula ing ne wo ks
o use hem in CI o DTNs is challenging. This pape analyzes
he possibili ies o ex ending he ne wo k emula o Minine
o p o ide high, de e minis ic usable bandwid h and de e -
minis ic, low ail la ency. Minine uses Linux namespaces o
isola e di e en nodes om each o he and e hs o p o ide
i ual links on he cu en node. Due o sha ed usage o
memo y, sys em buses, and CPU, mul iple po en ial non-
ne wo k-induced bo lenecks esul .
By ex ending Minine o suppo VMs as nodes and SR-
IOV-based links, we show ha signi ican imp o emen o
bo h bandwid h and la ency is possible simul aneously. Wi h
ou measu emen s on a se en-node line opology using an
ex e nal imes ampe and load gene a o , we show ha using
ha dwa e-suppo ed in e aces and cabled connec ions im-
p o es he highes ze o loss packe a e om 270 kpacke s/s
o 850 kpacke s/s. I p o ides as well a signi ican ly lowe
and mo e s able la encies. To conclude, using ou p o o ype
o e i y and es la ency-sensi i e o high h oughpu appli-
ca ions, se ices, o p o ocols is possible using he s anda d
1h ps://gi hub.com/ umi8/CIMinine
2h ps:// umi8.gi hub.io/minine - m-s io
3h ps://doi.o g/10.14459/2025mp1773238
Minine API. This allows es ing sc ip s and au oma ing he
in eg a ion o ne wo k applica ions in CI p ocesses o imp o e
he quali y o so wa e, se ices, and ools in ne wo ks.
Fu he , we plan o analyze he in luence o complex sce-
na ios, compu a ion-hea y applica ions, and Docke con aine s
as nodes oge he wi h SR-IOV in he u u e o imp o e he
possibili ies o ne wo k emula ion o la ency and h oughpu -
sensi i e applica ions. We u he plan o use ou published
da a o p o ide o e ime modeling and analyze ou da a’s
p edic abili y.
ACKNOWLEDGMENTS
This wo k was suppo ed by he EU’s Ho izon 2020 p o-
g amme as pa o he p ojec s SLICES-PP (10107977) and
G eenDIGIT (4101131207), by he Ge man Fede al Minis y
o Educa ion and Resea ch (BMBF) unde he p ojec s 6G-
li e (16KISK002) and 6G-ANNA (16KISK107), and by he
Ge man Resea ch Founda ion (Hype NIC, CA595/13-1).
REFERENCES
[1] M. Fowle and M. Foemmel, “Con inuous in eg a ion,” 2006.
[2] Y. Wu, K. Zhang, and Y. Zhang, “Digi al win ne wo ks: A su ey,” IEEE
In e ne o Things Jou nal, ol. 8, no. 18, pp. 13 789–13 804, 2021.
[3] P. Almasan, M. Fe iol-Galmes, J. Paillisse, and J. Su´
a ez-Va ela,
“Digi al Twin Ne wo k: Oppo uni ies and Challenges,” Jan 2022.
[Online]. A ailable: h ps://a xi .o g/pd /2201.01144
[4] A. Mozo, A. Ka amchandani, S. G´
omez-Cana al, M. Sanz, J. I. Mo eno,
and A. Pas o , “B5gemini: AI-D i en Ne wo k Digi al Twin,” Senso s,
ol. 22, no. 11, p. 4106, 2022.
[5] H. Wang, Y. Wu, G. Min, and W. Miao, “A g aph neu al ne wo k-based
digi al win o ne wo k slicing managemen ,” IEEE T ansac ions on
Indus ial In o ma ics, ol. 18, no. 2, p. 1367–1376, 2022.
[6] M. Sa a anan, P. S. Kuma , and A. R. Kuma , “Enabling ne wo k digi al
win o imp o e QoS pe o mance in communica ion ne wo ks,” IEEE
T ansac ions on Indus ial In o ma ics, 2022.
[7] G. Elec ic, “The Digi al Twin: Comp essing Time o Value o Digi al
Indus ial Companies,” Gene al Elec ic, 2018.
[8] R. S¨
ode be g, K. W¨
a me jo d, J. S. Ca lson, and L. Lindk is , “Towa d
a Digi al Twin o eal- ime geome y assu ance in indi idualized
p oduc ion,” CIRP annals, ol. 66, no. 1, pp. 137–140, 2017.
[9] U. Muslim and S. Recke , “Emula ion Pla o m o Build Digi al Twins
o Edge Compu ing En i onmen s,” in 2024 IEEE/ACM Symposium on
Edge Compu ing (SEC). IEEE, 2024, pp. 512–514.
[10] Y. Zheng, S. Yang, and H. Cheng, “An applica ion amewo k o digi al
win and i s case s udy,” Jou nal o Ambien In elligence and Humanized
Compu ing, ol. 10, pp. 1141–1153, 2019.
[11] F. Wiedne , M. Helm, S. Gallenm¨
ulle , and G. Ca le, “HVNe :
Ha dwa e-Assis ed Vi ual Ne wo king on a Single Physical Hos ,” in
IEEE INFOCOM 2022 - IEEE Con e ence on Compu e Communica-
ions Wo kshops, INFOCOM 2022 - Wo kshops, New Yo k, NY, USA,
May 2-5, 2022. IEEE, 2022, pp. 1–6.
[12] R. M. Fujimo o, K. S. Pe umalla, and G. F. Riley, Ne wo k simula ion.
Sp inge Na u e, 2022.
[13] A. Va ga, “OMNeT++,” in Modeling and ools o ne wo k simula ion.
Sp inge , 2010, pp. 35–59.
[14] A. K. Saluja, S. A. Da gad, and K. Mis y, “A De ailed Analogy o
Ne wo k Simula o s—NS1, NS2, NS3 and NS4,” In . J. Fu u e Re olu .
Compu . Sci. Commun. Eng, ol. 3, pp. 291–295, 2017.
[15] K. Kau , J. Singh, and N. S. Ghumman, “Minine as so wa e de ined
ne wo king es ing pla o m,” in In e na ional con e ence on communi-
ca ion, compu ing & sys ems (ICCCS). IEEE, 2014, pp. 139–42.
[16] J. Yan and D. Jin, “V -minine : Vi ual- ime-enabled minine o scalable
and accu a e so wa e-de ine ne wo k emula ion,” in P oceedings o
he 1s ACM SIGCOMM Symposium on So wa e De ined Ne wo king
Resea ch, 2015, pp. 1–7.
[17] M. Peus e , H. Ka l, and S. an Rossem, “MeDICINE: Rapid p o o yping
o p oduc ion- eady ne wo k se ices in mul i-PoP en i onmen s,” in
2016 IEEE Con e ence on Ne wo k Func ion Vi ualiza ion and So wa e
De ined Ne wo ks (NFV-SDN), No 2016, pp. 148–153.
[18] M. Peus e , J. Kampmeye , and H. Ka l, “Con aine ne 2.0: A Rapid
P o o yping Pla o m o Hyb id Se ice Func ion Chains,” in 2018 4 h
IEEE Con e ence on Ne wo k So wa iza ion and Wo kshops (Ne So ),
2018, pp. 335–337.
[19] Y. Dong, X. Yang, X. Li, J. Li, K. Tian, and H. Guan, “High pe o mance
ne wo k i ualiza ion wi h SR-IOV,” in HPCA - 16 2010 The Six een h
In e na ional Symposium on High-Pe o mance Compu e A chi ec u e,
2010, pp. 1–10.
[20] J. Liu, “E alua ing s anda d-based sel - i ualizing de ices: A pe o -
mance s udy on 10 GbE NICs wi h SR-IOV suppo ,” in 2010 IEEE
In e na ional Symposium on Pa allel Dis ibu ed P ocessing (IPDPS),
Ap . 2010, pp. 1–12.
[21] Z. Tang, D. Chen, T. Sun, L. Zhang, M. Qi, and X. Wang, “In elligen
Awa eness o Delay-Sensi i e In e ne T a ic in Digi al Twin Ne wo k,”
IEEE Jou nal o Radio F equency Iden i ica ion, ol. 6, pp. 891–895,
2022.
[22] E. Soa es, G. Sizilio, J. San os, D. A. Da Cos a, and U. Kulesza, “The
e ec s o con inuous in eg a ion on so wa e de elopmen : a sys ema ic
li e a u e e iew,” Empi ical So wa e Enginee ing, ol. 27, no. 3, p. 78,
2022.
[23] M. Vucnik, T. Solc, U. G ego c, A. H o a , K. B ega , M. Smolnika ,
M. Moho cic, and C. Fo una, “Con inuous In eg a ion in Wi eless
Technology De elopmen ,” IEEE Communica ions Magazine, ol. 56,
no. 12, pp. 74–81, 2018.
[24] S. Riege , L.-N. Lux, J. Schmi , and M. S ieme ling, “DigSiNe : Using
Mul iple Digi al Twins o P o ide Rhy hmic Ne wo k Consis ency,” in
NOMS 2024-2024 IEEE Ne wo k Ope a ions and Managemen Sympo-
sium, 2024, pp. 1–5.
[25] Z. Li, M. Kihl, Q. Lu, and J. A. Ande sson, “Pe o mance O e head
Compa ison be ween Hype iso and Con aine Based Vi ualiza ion,”
in 2017 IEEE 31s In e na ional Con e ence on Ad anced In o ma ion
Ne wo king and Applica ions (AINA), 2017, pp. 955–962.
[26] F. Wiedne , M. Helm, A. Daichend , J. And e, and G. Ca le, “Pe o -
mance e alua ion o con aine s o low-la ency packe p ocessing in
i ualized ne wo k en i onmen s,” Pe o m. E al., ol. 166, no. C, Jan.
2025.
[27] S. Gallenm¨
ulle , J. Naab, I. Adam, and G. Ca le, “5G QoS: Impac o
Secu i y Func ions on La ency,” in NOMS 2020 - IEEE/IFIP Ne wo k
Ope a ions and Managemen Symposium, Budapes , Hunga y, Ap il 20-
24, 2020. IEEE, 2020, pp. 1–9.
[28] T. Zhang, L. Linguaglossa, J. Robe s, L. Iannone, M. Gallo, and
P. Giaccone, “A benchma king me hodology o e alua ing so wa e
swi ch pe o mance o NFV,” in 2019 IEEE Con e ence on Ne wo k
So wa iza ion (Ne So ), Jun. 2019, pp. 251–253.
[29] G. K. Lockwood, M. Ta ineni, and R. Wagne , “SR-IOV: Pe o mance
Bene i s o Vi ualized In e connec s,” in Annual Con e ence o he
Ex eme Science and Enginee ing Disco e y En i onmen , XSEDE ’14,
A lan a, GA, USA - July 13 - 18, 2014, S. A. La h op and J. Alameda,
Eds. ACM, 2014, pp. 47:1–47:7.
[30] A. Daichend , F. Wiedne , J. And e, and G. Ca le, “Applicabili y o
Ha dwa e-Suppo ed Con aine s in Low-La ency Ne wo king,” in 2024
20 h In e na ional Con e ence on Ne wo k and Se ice Managemen
(CNSM). IEEE, 2024, pp. 1–7.
[31] AMD, “Pe o mance Tuning Guidelines o Low La ency Response
on AMD EPYC-Based Se e s Applica ion No e,” Jun. 2018, URL:
h ps://www.amd.com/con en /dam/amd/en/documen s/epyc- echnical-
docs/ uning-guides/56263-EPYC-pe o mance- uning-app-no e.pd ,
Las accessed: Jan 10, 2025.
[32] In el Co po a ion, “E810 da ashee e 2.5,” URL: h ps://cd d 2-
public.in el.com/613875/613875 E810 Da ashee Re 2.5.pd , Las Ac-
cessed: Jan 10, 2025.
[33] P. Emme ich, S. Gallenm¨
ulle , D. Raume , F. Wohl a , and G. Ca le,
“MoonGen: A Sc ip able High-Speed Packe Gene a o ,” in P oceedings
o he 2015 ACM In e ne Measu emen Con e ence, IMC 2015, Tokyo,
Japan, Oc obe 28-30, 2015, K. Cho, K. Fukuda, V. S. Pai, and
N. Sp ing, Eds. ACM, 2015, pp. 275–287.
[34] S. Gallenm¨
ulle , F. Wiedne , J. Naab, and G. Ca le, “Ducked ails:
imming he ail la ency o ( ) packe p ocessing sys ems,” in 2021 17 h
In e na ional Con e ence on Ne wo k and Se ice Managemen (CNSM).
IEEE, 2021, pp. 537–543.
